Dispatch-dependent dual-queue planning

⚠️ Curated Limitations & Negative Results

• The background calls D_i a deadline CDF and Eq.1 uses 1−D_i, but the deterministic-deadline paragraph then assigns D_i=1 before or on latest_start and 0 afterward. That decreasing step is inconsistent with the stated CDF role. Both inspected PDF versions retain the presentation conflict; no implementation defect, original-paper erratum or silently repaired convention is asserted.
• Eq.3 is a success probability, whereas Section 4.1 describes a failure score. The complementary interpretation is pf=Pr[m>d−t], with deadline equality counted as success; replacing > by >= needs an additional no-atom convention. The background Eq.1 survival convention and Eq.3 inclusive event must not be silently unified, and the prose does not provide a complete numerical implementation.
• For a fixed remaining-search law M and fixed deadline d, V(t)=Pr[m<=d−t] is nonincreasing. For nonnegative m and finite d, V(d)=Pr[m=0] and V(t)=0 is guaranteed for t>d (it can also be zero earlier); it need not be zero exactly at d. These fixed-model facts do not make the periodically refreshed scores monotonic when estimates, relaxed plans or deadlines change.
• Remaining search time is estimated from expansion-delay/search statistics and relaxed-plan length, while relaxed-plan deadline estimates are explicitly described as inadmissible. The scores are modeled probabilities, without a demonstrated calibration or error bound. The weighted f form does not transfer a weighted-A* quality bound, completeness theorem or successful-dispatch guarantee to this combined procedure.
• The none/all prefix argument requires a reasonable ancestor relaxed plan and a forward successor with a matching suffix. It matches assumed commitment contexts; it does not establish equality of node-specific search-time laws or probabilities, preservation of all relevant ancestors in the frontier, or sufficiency for arbitrary dispatch policies. Actual irreversible dispatch can make a goal unreachable.
• The abstract metareasoning background uses mutually exclusive non-preemptible domain actions with computation concurrent with at most one action. The selected temporal-planning setting and multi-robot evaluation allow concurrent actions. The abstract exclusivity restriction must not be assigned to every evaluated action, nor does that evaluation establish arbitrary uncertain execution dynamics.
• The paper gives prose queue and dispatch integration, not a complete executable specification. Duplicate-state handling, tie/cycle corner cases, sound dispatch and scheduling depend on the underlying planner and earlier dispatch procedure. Code is described as available from the authors upon request, but no request, code acquisition or algorithm execution occurred in this research.
• Section 4.1 associates OPTIC with Benton et al. (2012), while Section 5 cites Coles et al. (2012) for the implementation and the bibliography names that work COLIN. The discrepancy is retained without treating OPTIC and COLIN as aliases or diagnosing an implementation error.
• The evaluation sets w=5, uses the described 100-expansion score refresh and tests dispatch thresholds 0.025, 0.1 and 0.25. Removing subtree focus or the earlier allocate-t_u search restriction does not remove these parameters. No universally best threshold or expansion-rate-independent setting is established.
• The reported evaluation uses 100 RCLL scenarios with one-, two- and three-robot variants and eight simulated expansion rates from 10 to 1000 expansions per second. Those rates are simulated timing, not measured hardware speed or complete queue/heuristic overhead. Sums across rates reuse scenarios and are not counts of independent scenarios; the experiment does not establish general deployment performance.
• The paper’s virtual best solver selects a configuration by robot-count/domain and expansion-rate setting, not an explicitly instance-wise oracle. Aggregate coverage advantages are not uniform across settings, and the displayed tables report no confidence intervals or repeated-run uncertainty. No causal benefit or universal runtime advantage follows from the queue-agreement diagnostic.
• The author and official PDFs are distinct byte artifacts of one 2025 work, with different title-block forms of Shimony’s name and layout metadata. The institutional 2029 date remains a metadata question. The curated descriptive name disambiguates the source-local dual label from the maintained 1987 game-tree DUAL* algorithm; aliases are empty. All original bibliography and follow-up questions remain retained, not resolved by this draft.